基于改进YOLOv7网络的织物疵点检测  

Fabric defect detection based on improved YOLOv7 network

作　　者：石玉文 林富生 宋志峰 余联庆 SHI Yuwen;LIN Fusheng;SONG Zhifeng;YU Lianqing(College of Mechanical Engineering and Automation,Wuhan Textile University,Wuhan,Hubei 430200,China;Three Dimensional Textile Engineering Research Center of Hubei Province,Wuhan,Hubei 430200,China;Hubei Key Laboratory of Digital Textile Equipment,Wuhan,Hubei 430200,China)

机构地区：[1]武汉纺织大学机械工程与自动化学院,湖北武汉430200 [2]三维纺织湖北省工程研究中心,湖北武汉430200 [3]湖北省数字化纺织装备重点实验室,湖北武汉430200

出　　处：《毛纺科技》2024年第8期103-110,共8页Wool Textile Journal

摘　　要：针对传统的目标检测方法不能平衡检测精度、预测速度和轻量级部署模型、实时疵点检测,提出一种改进YOLOv7网络的轻量化检测模型。首先在主干网络引入轻量级卷积Ghost conv,在保证检测精度的同时降低网络参数量,提高对织物疵点的检测效率;其次添加CBAM注意力机制抑制无用信息,增强特征提取能力;最后在回归损失函数处引入新度量方法α-SIoU替换IoU,加速损失函数的自由度,提高网络模型的精确度。实验表明:该检测模型的准确率P达到96.27%,平均精度mAP值为83.84%,模型大小仅为19.10 MB,有效平衡了疵点检测的准确性、实时性与轻量级部署的问题。Aiming at the failure of traditional target detection methods to balance detection accuracy,prediction speed and lightweight deployment model to realize real-time defects detection,a lightweight detection model based on improved YOLOv7 network was proposed.Firstly,lightweight convolution Ghost conv was introduced into the backbone network to reduce the number of network parameters while ensuring the detection accuracy and improve the detection efficiency of fabric defects.Secondly,CBAM attention mechanism was added to suppress useless information and enhance feature extraction ability.Finally,a new measurement methodα-SIoU was introduced to replace IoU in the regression loss function to accelerate the degree of freedom of the loss function and improve the accuracy of the network model.The experiment shows that the accuracy P of the detection model reaches 96.27%,the average precision mAP value is 83.84%,the model size is only 19.10 MB,which effectively balances the accuracy,real-time and lightweight deployment of defects detection.

关 键 词：疵点检测 YOLOv7 轻量化 CBAM注意力机制 Ghost conv卷积 α-SIoU 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]